Over the last ten years we have witnessed a shift from large mainframe 
computing to commodity, off-the-shelf clusters of servers. Today's data 
centers contain thousands or tens of thousands of servers, providing 
services and computation for tens or hundreds of thousands of users. In 
addition to traditional IT challenges such as server management, security, 
and performance, data center owners now must deal with power and thermal 
issues, previously the domain of facilities management.  These trends will 
continue to accelerate as organizations acquire bladed servers and 
consolidate multiple, smaller clusters into centrally-located data 
centers. However, in spite of these trends, there has been no 
corresponding change in emphasis in the methods and toolkits that target 
system instrumentation, analysis, management, replay, and emulation.

This paper seeks to address this gap. We focus on methods and toolkits to 
enable the automated collection and analysis of workload traces from data 
centers, and use those traces as the basis for repeatable and verifiable 
experiments and workload emulation.  Our work has two components:

~ a location- and environment-aware extended knowledge plane that places 
thermal and power management concerns at the same level as service 
performance, collecting and analyzing facilities and performance data with 
particular focus on causal relationships across this boundary, and

~ data analysis and and workload playback methods that allow detailed and 
flexible emulation of enterprise-class workloads.

We discuss the high-level architectural requirements for these two 
components and present results from specific implementations and toolkits.